Separating the Effects of Risk Attitudes, Cognitive and - - PowerPoint PPT Presentation

Introduction Overview of Literature Data Econometric Model Very Preliminary Results

Separating the Effects of Risk Attitudes, Cognitive and Non-Cognitive Skills, Trust and Financial Aid in Schooling Decisions

Mathias ANDRÉ (École Polytechnique, CREST), Christian BELZIL (École Polytechnique, ENSAE, IZA), François POINAS (Toulouse School of Economics), Konstantinos TATSIRAMOS (IZA) Internal Seminar, CREST-LMI January 12, 2011

Internal Seminar CREST-LMI Separating Effects in Schooling Decisions

Introduction Overview of Literature Data Econometric Model Very Preliminary Results

Outline

1 Introduction 2 Overview of Literature 3 Data

German Socio Economic Panel and Background Variables Psychometric questionnaire

4 Econometric Model

Factors Model à la Cunha & Heckman Likelihood and Identification

5 Very Preliminary Results

Internal Seminar CREST-LMI Separating Effects in Schooling Decisions

Introduction Overview of Literature Data Econometric Model Very Preliminary Results Context and Framework

Motivation

Usual determinants of higher education : cognitive skills, parental backgrounds and other covariates, liquidity constraints. Deeper in the individual usually unobserved heterogeneity (caracteristics or preferences) :

Non cognitive skills and personality traits (or social preferences), Risk aversion, risk perception Social preferences, early childhood investments, etc.

Difficulty to have direct measures in practice (LC unmeasured by definition) and no unified framework.

Internal Seminar CREST-LMI Separating Effects in Schooling Decisions

Introduction Overview of Literature Data Econometric Model Very Preliminary Results Context and Framework

Remarks on Risk Influence

Risky accumulation process of human capital : uncertainties on return to schooling Risks of success/failure (imperfect information set) Uncertainty of finding a job after graduating, about the real quality of diploma (wage distribution unknown) Education can also be an insurance against risks on the labor market and have an effect on risk exposure (income fluctuations, unemployement, etc.). Imperfections on loan markets (no consensus about returns to schooling) can cause liquidity constraints and uncover the risks. Can risk aversion influence schooling choices ?

Internal Seminar CREST-LMI Separating Effects in Schooling Decisions

Introduction Overview of Literature Data Econometric Model Very Preliminary Results Context and Framework

Framework (1/2)

We use German Socio Economic Panel (GSOEP) between 2000 and 2008 with personality questionnaire, risk attitudes and background variables. We have scolarity choices and labor income. Our approach is tied to both Psychometrics (self-reported measurements) and to Economics (real life economic

• utcomes, not only tests).

German schooling system is particular with two main tracks : academic (Gymnasium) and professional. It is called early tracking because children are selected at age 11.

Internal Seminar CREST-LMI Separating Effects in Schooling Decisions

Introduction Overview of Literature Data Econometric Model Very Preliminary Results Context and Framework

Framework (2/2)

With a simple behavioral model, we study different choices :

Occupational choices for professionnal track (P) Drop out (G1) and continuation probability to university (G2) for academic.

Application of Factors Models as Carneiro, Hansen, Heckman (IER 2003) by maximum likelihood with measurement errors and a simulated random part for factors. We will use Heckman & Singer’s framework for true unobserved heterogeneity for the financial aid variable. Two parts in the likelihood : Extraction of the distribution of Factors and Estimation of the discrete choice model.

Internal Seminar CREST-LMI Separating Effects in Schooling Decisions

Introduction Overview of Literature Data Econometric Model Very Preliminary Results Context and Framework

Outline

1 Introduction 2 Overview of Literature 3 Data

German Socio Economic Panel and Background Variables Psychometric questionnaire

4 Econometric Model

Factors Model à la Cunha & Heckman Likelihood and Identification

5 Very Preliminary Results

Internal Seminar CREST-LMI Separating Effects in Schooling Decisions

Introduction Overview of Literature Data Econometric Model Very Preliminary Results

Outline

1 Introduction 2 Overview of Literature 3 Data

German Socio Economic Panel and Background Variables Psychometric questionnaire

4 Econometric Model

Factors Model à la Cunha & Heckman Likelihood and Identification

5 Very Preliminary Results

Internal Seminar CREST-LMI Separating Effects in Schooling Decisions

Introduction Overview of Literature Data Econometric Model Very Preliminary Results Related works

Factors Models and Noncognitive Skills

Cunha, Heckman (AER 2007, JHR 2008), Heckman, Stixrud, Urzua (JLE 2006) Cunha, Heckman, Schennach (Ecta 2010) :

Non linear dynamic factors (CS and NCS with endogenous parental investments) non parametrically identified (linearly estimated on CNLSY) with additive measurement errors terms Importance of measurement errors, of timing of investments,

• f self-productivity effects (only cross for NCS on CS)

IZA Workshop (01/2011): Cognitive and Non-Cognitive Skills

Internal Seminar CREST-LMI Separating Effects in Schooling Decisions

Introduction Overview of Literature Data Econometric Model Very Preliminary Results Related works

Risk Aversion and Schooling Decisions

Lehvari, Weiss (AER 1974), Belzil, Hansen (2004), Brodaty, Gary-Bobo, Prieto (2010) Special Issue of Labour Economics (2007), Sociological litterature (Holm, Jäger (2006)) Belzil, Leonardi (2007) :

Given high school completion, continuing to higher education negatively correlated with risk aversion. Individual differences in risk aversion almost as important as parents’ educational background and with good explanatory power (e.g. resp. 30% against 55% of variance) But no direct measure of heterogeneity.

Dohmen et al. (2005) : good quality of the GSOEP.

Internal Seminar CREST-LMI Separating Effects in Schooling Decisions

Introduction Overview of Literature Data Econometric Model Very Preliminary Results Related works

Behavorial models and social preferences

Bohnet, Zeckhauser (2004), Fehr (2009) : Role of trust in social and market transactions (beliefs) Behavioral and statistical evidence that trust is not just a special case of risk taking (using GSOEP) If risky context, trust can be a critical ingredient Imperfect measure and need of link it with risk behavior

Internal Seminar CREST-LMI Separating Effects in Schooling Decisions

Introduction Overview of Literature Data Econometric Model Very Preliminary Results Related works

The Liquidity constraints question

Leslie (1984), Keane (AER 2002) Keane, Wolpin (IER 2001) :

Methodological contribution (simulated likelihood with classification errors, form of EMAX, terminal period). Structural estimation (dynamic school/work model with savings decisions). LC exist but no impact on enrollment, substantial parental transfers increase attendance.

Internal Seminar CREST-LMI Separating Effects in Schooling Decisions

Introduction Overview of Literature Data Econometric Model Very Preliminary Results

Outline

1 Introduction 2 Overview of Literature 3 Data

German Socio Economic Panel and Background Variables Psychometric questionnaire

4 Econometric Model

Factors Model à la Cunha & Heckman Likelihood and Identification

5 Very Preliminary Results

Internal Seminar CREST-LMI Separating Effects in Schooling Decisions

Introduction Overview of Literature Data Econometric Model Very Preliminary Results German Socio Economic Panel and Background Variables

German schooling system and Choices diagram

Earling tracking (age 11) : Professionnal (Apprenticeship, Vocational) versus Academic (Gymnasium). If Abitur is completed, one can enter higher education (University).

Initial Condition Academic Track AP G1 G1 V G1 W I G1 Abitur G2 AP V W I U Professional Track AP P V P W P I P Internal Seminar CREST-LMI Separating Effects in Schooling Decisions

Introduction Overview of Literature Data Econometric Model Very Preliminary Results German Socio Economic Panel and Background Variables

Covariates

The 10 Covariates are : Indicators of being in the Gymnasium track (at age 17), gender, indicator if individual is born in former East Germany Indicators if father and mother has attended Gymnasium (cross product), Household size, Number of siblings between 0-14 when individual is aged 17 Yearly per capita income and Government-provided financial aid for higher education (in thousands of 2005 euros) This Financial Aid Indicator is very useful (see below).

Internal Seminar CREST-LMI Separating Effects in Schooling Decisions

Introduction Overview of Literature Data Econometric Model Very Preliminary Results German Socio Economic Panel and Background Variables

Financial Aid to go to University

It is a means-tested transfer scheme depending on parental income, taxes and social contributions. Half of the financial aid is paid as a grant and the other half as a loan that has to be repaid after completion of tertiary education. The maximum monthly amount of financial aid is 585 Euros (fully granted if monthly parents’ income less than 1,440 Euro). The part of parents’ income exceeding the threshold is withdrawn at a rate of 50% : Grant = 585 if Inc ≤ 1, 440 Max(0; 585 − (Inc − 1, 440)/2) if Inc > 1, 440

Internal Seminar CREST-LMI Separating Effects in Schooling Decisions

Introduction Overview of Literature Data Econometric Model Very Preliminary Results German Socio Economic Panel and Background Variables

Descriptive Statistics

Variables/Moda 1 Freq. Percent. Female 1238 49.6 Mother Educ. 374 15.2 Father Educ. 442 18.2 Born in East G. 686 27.5

Table: Binary Covariates

Note : Education is indicator for finishing Gymnasium Variables Mean

• Std. Dev.

Max Household 4.1 1.23 13 Siblings 0.5 0.83 7 Income 10.5 6.13 99.03 Grant 2.44 2.91 7.0

Table: Other Covariates

Note : Money amount are yearly and in ’000 euros 2005.

Internal Seminar CREST-LMI Separating Effects in Schooling Decisions

Introduction Overview of Literature Data Econometric Model Very Preliminary Results Psychometric questionnaire

Measuring Attitudes and Skills (1/2)

Cognitive skills (3q) : Score based on the IST-2000R test developed by Amthauer et al. (2001) with 3 components : verbal, numerical, and figural abilities. Non cognitive skills and Social preferences (10q+3q): Personality questionnaire based on the Big Five dimensions from Psychology; Trust on other people, society etc. (motivation, social preferences, values, perception of outside world, etc.) General risk question (1q) : "How do you see yourself: Are you generally a person who is fully prepared to take risks or do you try to avoid taking risks?"

Internal Seminar CREST-LMI Separating Effects in Schooling Decisions

Introduction Overview of Literature Data Econometric Model Very Preliminary Results Psychometric questionnaire

Measuring Attitudes and Skills (2/2)

Risk attitudes (6q) : Same question but in specific context (car driving, financial matters, leisure and sports, career, and health) Lottery (1q) : "Imagine you had won 100,000 Euros in a lottery. Almost immediately after you collect, you receive the following financial

• ffer from a reputable bank, the conditions of which are as follows:

There is the chance to double the money within two years. It is equally possible that you could lose half of the amount invested." − → K = 24 measurement questions

Internal Seminar CREST-LMI Separating Effects in Schooling Decisions

Introduction Overview of Literature Data Econometric Model Very Preliminary Results Psychometric questionnaire

Structure of the Sample (1/2)

Timing of the model is :

1 At age 11, children are distributed in Tracks : Professionnal

(P) or Academic (G). Everything will be conditional on it.

2 At age 17 (or around for University), they have a distribution

for factors and make choices according to them and their X’s.

We have realized choices for ??? individuals. We only estimate factors distributions on 17-year old individuals (between ?? and ?? non missing questions). Three types of individuals in the data :

Only schooling choices ?? Only psychometric questionnaire for factors ?? In both parts of the likelihood ??

Internal Seminar CREST-LMI Separating Effects in Schooling Decisions

Introduction Overview of Literature Data Econometric Model Very Preliminary Results Psychometric questionnaire

Structure of the Sample (2/2)

Missing data are supposed random, years of interview are :

Data \ Year 00-03 2004 2005 2006 2007 2008 Questions type : lottery (1q)

• X
• risk taking (6q)
• X
• general risk (1q)
• X
• X

X

• non cognitive (10q)
• X

X X

• cognitive (3q)
• X
• trust (3q)
• X

X

• Choices

X X X X X X Sample size 1278 328 315 269 305 .

Note: Entries marked by an ”X ” denote a year over which a specific set of questions has been administered.

Internal Seminar CREST-LMI Separating Effects in Schooling Decisions

Introduction Overview of Literature Data Econometric Model Very Preliminary Results Psychometric questionnaire

Composition of the Sample : Choices

Academic Status at age 17 Track Frequency Percentage Professional (P) 1467 58.8% Academic (G) 1028 41.2% Total 2495 100% Choices in Percentage by Group

Likelihood Choices Tree

Choices \ Group Professional (P) Academic (G1) If completed Abitur (G2) Apprenticeship 45% 9% 4% Vocational 17% 7% 1% Work 13% 3% 10% Inactivity 25% 6% 16% Complete Abitur

• 75% (G2)
• Higher Education
• 69%

Total 100% 100% 100%

Internal Seminar CREST-LMI Separating Effects in Schooling Decisions

Introduction Overview of Literature Data Econometric Model Very Preliminary Results

Outline

1 Introduction 2 Overview of Literature 3 Data

German Socio Economic Panel and Background Variables Psychometric questionnaire

4 Econometric Model

Factors Model à la Cunha & Heckman Likelihood and Identification

5 Very Preliminary Results

Internal Seminar CREST-LMI Separating Effects in Schooling Decisions

Introduction Overview of Literature Data Econometric Model Very Preliminary Results Factors Model à la Cunha & Heckman

Modeling Factors

We define 4 factors Fj : risk aversion (denoted RA), cognitive ability (denoted CO), non cognitive skills or motivation (denoted MO) and social preferences or trust (denoted TR). Each Fj is the sum of two orthogonal parts : a deterministic component and a random (unobserved to the econometrician) component which follows a Normal distribution : Fj = Fj + F ∗

j with F ∗ j ∼ N(0, σ2 Fj)

for 1 ≤ j ≤ 4 The deterministic part is linear in covariates and writes : Fj =

10

• r=1

φr

j · xr

where xr covariates and φr

j estimated parameters.

Internal Seminar CREST-LMI Separating Effects in Schooling Decisions

Introduction Overview of Literature Data Econometric Model Very Preliminary Results Factors Model à la Cunha & Heckman

Measurement Equations

Psychometric questions are discrete ordered variable and noted Mk, each driven by a continuous latent variable M∗

k

Latent variables are linear in factors : M∗

k = αk + 4

• j=1

αk

j · Fj

• m∗

k

+εk for k = 1, 2, ...K where εk ∼ N(0, σk) is the measurement error term. These measurements are extracted from four groups of questions (Risk, NCS, Social Preferences, tests) so that we impose some constraints on the α’s (see exclusion restrictions).

Internal Seminar CREST-LMI Separating Effects in Schooling Decisions

Introduction Overview of Literature Data Econometric Model Very Preliminary Results Factors Model à la Cunha & Heckman

Restrictions on Factors

Not a strictly statistial approach : limitation of number of factors (4) and exclusion constraints are imposed. Exclusion restrictions (on αk

j ’s) are :

type of question # of questions RA MO CO TR lottery 1 X . . . risk taking 6 X X X X general risk 1 X X X X non-cognitive 10 . X . X trust 3 . . . X cognitive 3 . X X .

Internal Seminar CREST-LMI Separating Effects in Schooling Decisions

Introduction Overview of Literature Data Econometric Model Very Preliminary Results Factors Model à la Cunha & Heckman

Extracting the Structural Risk Aversion

From the lottery question, six possible responses for the amount invested : sh = h

5 for h = 0, ..., 5 is the proportion of

x = 100, 000 invested in the lottery. Assuming CRRA utility function, payoff is : U(sh, x) = ((1 − sh)x)1−RAi 1 − RAi +0.5(2shx)1−RAi 1 − RAi +0.5(0.5shx)1−RAi 1 − RAi With an additional randorm term N(0, σL) (expected utility unobserved to the econometrician ex ante) and known fixed thresholds th, the likelihood of answering sh ∈ [th, th+1] is : Φ(U(th+1, x) σL ) − Φ(U(th, x) σL )

Internal Seminar CREST-LMI Separating Effects in Schooling Decisions

Introduction Overview of Literature Data Econometric Model Very Preliminary Results Likelihood and Identification

Behavioral Model : logit specification

Occupational Choices :

1 Professional Track (group P) : k ∈ CP1 = {AP, V , W , I} 2 Gymnasium Track (group G1) : k ∈ CG1 = {AP, V , W , I, AB}

Continuation Probability (if Abitur, group G2) : k ∈ CG2 = {AP, V , W , I, HE} Logit specification for choice k in group g ∈ {P, G1, G2}: P(k | i ∈ g) = exp(Uk,i∈g)

• k′∈Cg exp(Uk′,i∈g) for k ∈ Cg

where Uk,i∈g = α0kg + αRA,k,g · RAi + αMO,k,g · MOi + αCO,k,g · COi + αTR,k,g · TRi + αX,k,g · Xi

Internal Seminar CREST-LMI Separating Effects in Schooling Decisions

Introduction Overview of Literature Data Econometric Model Very Preliminary Results Likelihood and Identification

Two Parts of the Likelihood

Choices ci of individual i in group g (P, G1 or G2) :

Choices

n

• i=1
• 1≤g≤3

groups

• k∈Cg
• choices

P(k|i ∈ g)1(ci=k) with P(ci = k|i ∈ g) = eUk,i∈g

• k′∈Cg

eUk′,i∈g Measurements mk,i of individual i of question 1 ≤ k ≤ K (nk modalities)

n

• i=1

K

• k=1
• measures

nk

• j=1
• modalities

P(Mk,j)1(mk,i=Mk,j) with P(mk,i = Mj,k) = Φ(αj+1,k − m∗

k,i) − Φ(αj,k − m∗ k,i)

Internal Seminar CREST-LMI Separating Effects in Schooling Decisions

Introduction Overview of Literature Data Econometric Model Very Preliminary Results Likelihood and Identification

Identification

Theorems of Heckman and coauthors (2003, 2010) : non parametric identification of non linear factors models with endogenous inputs With constraints on parameters :

Identification for linear factors model : αk = 0, σk = 1 for 1 ≤ k ≤ K and φi = 0 for 1 ≤ i ≤ 4 (resp. intercept and standard errors in measurement equations and intercept in deterministic part of factors). Because we identify all thresholds in probit models of measurement equations. Identification of logit models in groups P, G1 and G2 : αi,I = 0 and βk,i,I = 0 for factor 1 ≤ i ≤ 4 and choice 1 ≤ k ≤ 2 (I - inactivity- is the reference modality)

Internal Seminar CREST-LMI Separating Effects in Schooling Decisions

Introduction Overview of Literature Data Econometric Model Very Preliminary Results

Outline

1 Introduction 2 Overview of Literature 3 Data

German Socio Economic Panel and Background Variables Psychometric questionnaire

4 Econometric Model

Factors Model à la Cunha & Heckman Likelihood and Identification

5 Very Preliminary Results

Internal Seminar CREST-LMI Separating Effects in Schooling Decisions

Introduction Overview of Literature Data Econometric Model Very Preliminary Results Estimation results

Estimates of Factors

Importance of Factors with respect to measurement errors :

σk

√

V(Fj) reported by type of measurement equations Ratio/Q Risk NCS

• Cog. Sk.

Trust <0.7 1 3 . 1 [0.7;1.3] 5 1 . 1 >1.3 2 6 3 1

Covariance between deterministic part of Factors :

RA MO CO TR RA 1506 . . . MO

• 1016

3016 . . CO

• 1403
• 187

2325 . TR

• 2424
• 704

3575 7484

Table: Covariances

Note : Sum of Variances is S = 14330.54 RA MO CO TR RA 1 . . . MO

• .48

1 . . CO

• .75
• .07

1 . TR

• .72
• .15

.12 1

Table: Correlation Matrix

Internal Seminar CREST-LMI Separating Effects in Schooling Decisions

Introduction Overview of Literature Data Econometric Model Very Preliminary Results Estimation results

Estimates of Probabilities

Effects on probabilities : Variable/Proba.

• Appr. (P)
• Appr. (G1)
• Univ. (G2)

Risk Aversion 0.53

• 8.89

5.66 Cognitive ability 6.87 0.68

• 9.91

Motivation

• 6.42
• 3.49

1.38 Trust (Social P.) 0.33

• 4.72

3.58 Financial Aid

• 5.06
• 1.86
• 3.43

Father’s Educ.

• 7.93
• 2.83
• 4.26

Standard errors still non reliable and may need to be bootstrapped We will report marginal effects and percent correctly predicted

Internal Seminar CREST-LMI Separating Effects in Schooling Decisions

Introduction Overview of Literature Data Econometric Model Very Preliminary Results Final Comments

Conclusion

Optimization of the likelihood is currently running in Fortran, waiting for final and more reliable results (691-214=417 estimated parameters). We implemented a Factors model of schooling choices taking into account various of dimensions : cognitive and non cognitive skills, risk aversion and liquidity constraints Extensions : with Gymnasium, personality may evolve (dynamic risk aversion) and additional unobserved heterogeneity in the Financial Aid variable.

Internal Seminar CREST-LMI Separating Effects in Schooling Decisions

Introduction Overview of Literature Data Econometric Model Very Preliminary Results Final Comments

The End

Thank you.

Internal Seminar CREST-LMI Separating Effects in Schooling Decisions

Introduction Overview of Literature Data Econometric Model Very Preliminary Results Final Comments

Some questions on personality

Do you agree with these statements ? (Pick boxes between "Totally agree" and "Totally disagree") NCS (10) : Life depends on me, I have not what I deserve, Achievements in life is fate or luck, Social activity can affect social conditions, Others influence life, Work hard to succeed, etc. Trust (3) : one can trust people, one can’t rely on anyone, better to be careful with strangers.

Internal Seminar CREST-LMI Separating Effects in Schooling Decisions